For a store that gets two negative reviews a week, five in a single day is a crisis; for a mall store that gets fifty reviews a week, it is an ordinary Tuesday. A fixed threshold cannot tell the two apart. Review anomaly detection compares every store and every topic with its own history and flags unexpected spikes without you writing a single rule. This post explains how BranchVoice finds anomalies, which types it watches, and what a regional manager should do with an open one.
What review anomaly detection does that thresholds cannot
An alert rule answers "tell me when this happens"; anomaly detection answers "tell me what changed, even if I never thought to ask". The two do not replace each other. Rules give you precision on known risks: e-mail me when the 1-star count passes 3. Anomalies catch what you did not define: a topic that was silent for three months suddenly dominating a week of reviews, or a store's average rating falling far below its last ninety days.
For a chain the practical value is not having to maintain a separate threshold per store. Five hundred stores have five hundred different "normals", and anomaly detection learns each one from the data.
How the baseline is built
BranchVoice evaluates the most recent week against the eight weeks before it. For each week it computes the relevant metric (review count, or a topic's share of analysed reviews), then compares the latest week with the mean and spread of the previous eight. When the deviation is statistically large, an anomaly opens.
A floor protects against small numbers: if the latest week has too few reviews (a topic that appears in only two reviews, say), it is not counted as a spike. Random wobble at low-volume stores therefore does not turn into an anomaly.
The check runs overnight and is re-triggered for a company after a large sync or a batch of AI analysis. For the same type, store and topic only one open record is kept for 14 days, so the same deviation does not spawn a new card every night.
The anomaly types that are tracked
| Type | Scope | What it catches |
|---|---|---|
| Volume spike | Brand | Weekly total review count far above the baseline |
| Topic spike | Brand | A topic's share of analysed reviews rising sharply |
| Rating drop | Store | Last 7 days' average rating at least 0.5 stars below the last 90 days |
| Suspicious burst | Brand | 3 or more reviews the AI flagged as suspicious within 7 days |
| Critical review | Review | A single review whose risk score reaches the top level (hygiene, safety, misleading pricing and the like) |
The topic spike is usually the earliest signal. When a campaign's terms are misunderstood or a product batch has a quality defect, the "price and discounts" or "durability" topic rises first on Ekşi Sözlük (a large Turkish community forum) and Şikayetvar (Turkey's main consumer complaint site), then in Google reviews. Because it is measured at brand level, reviews that could not be matched to a store are included; see AI review classification for how that works.
What to do once an anomaly appears
Open anomalies are listed on the "Anomalies" tab of the Alerts page and counted in the dashboard's "To do" card. Rating drops and suspicious bursts raise warning-level in-app notifications; volume and topic spikes raise info-level ones. Each card shows the numbers behind the deviation (current value, baseline), the store involved and the triggering reviews with their summaries.
A regional manager's routine has three steps:
- Verify. Read the reviews and work out whether the deviation reflects a real event (renovation, staff change, stock problem) or one angry customer posting several times.
- Own it. Move the anomaly to "acknowledged" so the rest of the team sees that someone is on it.
- Close and codify. Close the record once the issue is resolved. If the same deviation recurs, it is now a known risk; define a clear threshold for it with critical review alerts.
Suspicious bursts deserve separate attention. Positive reviews arriving in quick succession with similar wording were often collected at the store's own initiative and may breach platform rules; clusters of similar negative reviews may come from a competitor. In either case, inform head office before anyone replies.
Where anomaly detection fits in chain operations
Anomaly detection sets the agenda for the weekly meeting. Alerts from rules run the daily operation, while anomalies answer "what changed across the chain this week" with evidence. Three checks are enough: is the number of open anomalies growing, is the same store entering a rating drop repeatedly, and are topic spikes clustering in one region.
The same questions appear in the weekly digest e-mail and in scheduled reports, so anomalies reach the manager who reads the report as well as the person who logs into the app.
Next steps
Anomaly detection needs no setup: once your stores are connected and a few weeks of review history exist, the baseline forms on its own and deviations start being flagged. To see which deviations your own stores produce, request a demo.